Qwen3.5-9B — MatRL multi-turn SFT (cold start)
Qwen/Qwen3.5-9B cold-started on
amayuelas/matrl-sft-mt:
1,119 multi-turn tool-use episodes for crystal-structure inverse design.
This is the model-scale arm of
amayuelas/Qwen3.5-4B-MatRL-MT-SFT.
It uses the same corpus, renderer, 3-epoch schedule and loss mask. Only the
model, the parallelism layout and the learning rate differ, so any difference
between the two comes down to scale.
This is the cold-start checkpoint. Its job is format and the tool loop:
propose, evaluate, refine, and above all commit within the turn budget. The
chemistry is learned afterwards by RL; see
amayuelas/Qwen3.5-9B-MatRL-MT-RL.
Why a cold start is needed
Base Qwen3.5 essentially never commits. Across 2,880 multi-turn rollouts the 4B
base model made 3 submit calls. It proposes prolifically (~5.6
candidates/rollout), evaluates sparingly, then runs out of turns. Strict
multi-turn SUN is therefore 0% by construction: a commitment failure, not a
chemistry failure. This stage teaches the agent to close an episode.
Training
| base | Qwen/Qwen3.5-9B |
| data | 1,119 of the dataset's 1,123 episodes, assistant-only loss |
| steps | 420 (3 epochs), global batch 8 |
| seq len | 16,384 |
| optimizer | AdamW, lr 1e-5, constant, no warmup (half the 4B's 2e-5) |
| precision | bf16 |
| parallelism | FSDP over 8 ranks + context parallel (cp=4, ulysses) |
| trainer | prime-rl |
Trained on 8×A100-40GB. Loss fell from 1.21 at the first step to ~0.8 over the final epoch.
Four episodes (3 band-gap, 1 formula) are longer than 16,384 tokens once rendered with the Qwen3.5 chat template, so they were dropped instead of truncated.
Context parallelism uses ulysses, not ring: Qwen3.5 is a hybrid with linear attention (DeltaNet) layers, and ring attention is a softmax-attention algorithm that does not apply to them.
Evaluation
This checkpoint has not yet been scored on the SUN benchmark used for the 4B release.
Important: thinking channel
This model is trained with reasoning in the native thinking channel
(reasoning_content → <think>), preserved across tool calls. Do not
evaluate it with enable_thinking=false, which turns off the very behavior
this run trains.
Serving
Qwen3.5 is a VL-capable model class, so vLLM requires an image-processor config
even for text-only serving. preprocessor_config.json and
video_preprocessor_config.json are included; without them vLLM fails on load.